CanIRun.ai has drawn attention on Hacker News by tackling a simple question for local AI users: what models can a given computer actually run? The web-based tool opens in a browser, detects GPU and memory specifications through the WebGPU API, and then estimates whether specific LLMs are runnable and how fast they may perform.
Its output goes beyond a basic yes-or-no check. The site assigns grades from S to F and shows estimated inference speed in tokens per second. Model coverage spans from lightweight 0.8B-parameter systems to massive 1T-parameter MoE, or mixture-of-experts, models. The dataset cited by the article includes major local inference tools such as llama.cpp, Ollama, and LM Studio.
Useful for beginners, but hardware coverage and estimates are under scrutiny
The appeal is obvious. New users trying local LLM deployment often do not know where to start, and a browser tool that inspects hardware automatically lowers the barrier. Still, criticism surfaced quickly. The main complaints center on incomplete hardware listings and a gap between estimated and real-world performance.
According to the source article, several GPUs are missing from the current list, including RTX Pro 6000, RTX 5060 Ti 16GB, and various laptop GPUs. Apple silicon is included, but the listed memory ceiling stops at 192GB, while the article notes that M3 Ultra can actually support up to 512GB. For users comparing hardware before running local models, omissions like these can distort the result.
Accuracy is the other weak spot. Some users said models that worked on their own systems were marked as unsupported by the site. Those cases — where real machines run a model but the web tool says otherwise — appeared repeatedly in discussion threads and led some people to stop relying on the scores.
CLI alternative llmfit enters the discussion
Community members also pointed to llmfit as an alternative. Unlike a browser-based checker, llmfit is a command-line program that can call system tools directly, including nvidia-smi, to retrieve GPU information with greater precision. That made it more practical and more accurate in the view of some users.
At the same time, llmfit raised another issue. Some users were surprised by how accurately hardware could be identified without an explicit permission prompt. That fed into a broader discussion about browser fingerprinting and hardware privacy. If a web tool can read graphics card details through WebGPU API access, users want to know how that information might be handled.
Users want model quality guidance, not just a run check
The feedback goes beyond pure hardware detection. What users really want to know is which model offers the best quality while still running at an acceptable speed on their machine. Right now, CanIRun.ai mainly answers whether a model can run at all. It does not yet tell users whether the experience will actually be good enough.
The article says the community wants benchmark-based model quality scores added to hardware estimates. Other requested upgrades include CPU memory sharing strategies for GPUs with limited memory, support for KV cache offloading techniques, and fixes to how MoE models are evaluated. Some users also suggested that this kind of feature would make more sense if integrated directly into Ollama, letting people filter models from the command line based on local hardware.
Even with its limitations, CanIRun.ai still offers a fast starting point for people exploring local AI. It addresses a real need — checking whether a machine is suitable before downloading and running large models — but the current version still leaves room for improvement in coverage, accuracy, and privacy clarity.